shahzeen/spectra_dataset

SPECTRA — RASMD preprocessed scene info

Pose, depth and pair annotations for RASMD (RGB–SWIR driving scenes), prepared for training and evaluating cross-spectral image matchers.

This repository holds the annotations only. The images are not redistributed here.

1. Get the images

Download the RGB and SWIR images from the official RASMD page:

https://yonsei-stl.github.io/RASMD/

Lay them out so that the relative paths stored in the scene files resolve:

RASMD_DATASET/
  1_sunny/data_3_urban/croped_data/rgb/rgb_aligned_000000.jpg
  1_sunny/data_3_urban/croped_data/swir/swir_original_000000.jpg
  ...

2. Get these annotations

huggingface-cli download shahzeen/spectra_dataset RASMD_prep_sce.tar --repo-type dataset
tar xf RASMD_prep_sce.tar -C RASMD_DATASET/

giving:

RASMD_DATASET/prep_sce/
  train/<scene>.npy     35 scenes
  val/<scene>.npy        3 scenes
  test/<scene>.npz      18 scenes
  depth/<scene>/<frame>.jpg.h5   33,488 depth maps

Every path inside is relative to RASMD_DATASET/, so nothing needs rewriting.

3. Data card

A frame is one pixel-aligned RGB + SWIR capture, so each frame contributes two image files and one depth map. A pair is one unordered cross-spectral image pair.

Split	Scenes	Frames	Images	Pairs
train	35	26,525	53,050	322,628
val	3	1,891	3,782	14,974
test	18	5,072	10,144	46,948
total	56	33,488	66,976	384,550
test — evaluation subset	14	—	—	1,500

Pairs by weather:

Split	sunny	cloudy	rainy	snowy
train	198,372	114,188	4,057	6,011
test	25,034	16,702	885	4,327
test subset (1,500)	476	476	309	239

The 1,500-pair evaluation subset is the benchmark split: drawn from 14 of the 18 test scenes, geometry-filtered, and balanced across weather so no single condition dominates. Results reported in the SPECTRA paper use this subset.

Pair direction

train and val store every pair twice, once per direction — (RGB_i, SWIR_j) and (SWIR_j, RGB_i).

Fields

train / val are pickled dicts (.npy) with image_paths, depth_paths, intrinsics, poses, pairs, overlaps. In each scene the first N entries of image_paths are the RGB frames and the next N the SWIR frames; both halves share intrinsics, pose and depth, since the two modalities are pixel-aligned.

test are .npz with rgb_paths, swir_paths, depth_paths, intrinsics, cam2world, pair_idx0, pair_idx1, pair_overlaps.

Citation

Please cite RASMD when using these annotations, and follow the licence on the RASMD page for the images themselves.